Storage Performance Tuning by Workload

Storage Performance Tuning by Workload An architecture diagram generated by Archify. Database · Low latency · random I/O · Architecture component Database Low latency · random I/O Web server · Shared content · RWX · Architecture component Web server Shared content · RWX Data analytics · Large sequential reads · Architecture component Data analytics Large sequential reads Machine learning · LLM · Model load throughput · Architecture component Machine learning · LLM Model load throughput EBS gp3 · io2 · IOPS and throughput set apart · Architecture component EBS gp3 · io2 IOPS and throughput set apart EFS · Elastic throughput mode · Architecture component EFS Elastic throughput mode FSx for Lustre · Striping · compression · Architecture component FSx for Lustre Striping · compression IOPS-bound Shared across nodes Throughput-bound Large sequential

EBS tuning

  • • gp3 decouples IOPS and throughput from capacity, so you avoid buying size you do not need
  • • For hundreds of thousands of IOPS, use io2 Block Express on an EBS-optimized instance
  • • If the instance's EBS bandwidth ceiling is below the volume's, the instance is the bottleneck

EFS tuning

  • • Default to elastic throughput and switch to provisioned only when you need steady performance
  • • Many-small-file reads are latency-sensitive, so add a caching layer
  • • Tune read and write buffer sizes in the mount options to raise throughput

FSx and LLM workloads

  • • Decide deployment type and provisioned throughput (125-1,000 MB/s per TiB) first
  • • Striping spreads large files across storage targets to widen parallel reads
  • • Workloads that reload model weights repeatedly, such as vLLM, favor FSx for Lustre